[uml-devel] Re: rootstrap: a tool for building Debian root filesystems for UML

[Matt Z. <md...@de...>]

And no sooner do I announce this than it breaks...debootstrap 0.1.17 just
hit the archive, and rootstrap no longer works (though it does with
0.1.16.4).

It looks like yet another race that only occurs under UML (debootstrap
0.1.17 works fine outside of UML)

A very similar problem has always occurred when I tried to build potato
systems with rootstrap, but I ignored it because I don't build potato images
very often, and it's possible to work around the problem.  This new
incarnation, though, is very painful to work around.

The old symptom (debootstrap 0.1.16.4 and potato installation) was that the
installation would hang during the postinst of perl-5.005-base.  A copy of
that script is attached.  If I strace the hung process inside UML (a shell),
I get:

write(1, "\n", 1)                       = -1 EPIPE (Broken pipe)
--- SIGPIPE (Broken pipe) ---

At which point strace exits, but the process does not.  I can repeat this as
many times as I like with the same result.  If I send a SIGHUP to the
process, everything continues on, and installation is successful.  So I
never bothered tracking it down.

The new problem (debootstrap 0.1.17 and woody) manifests itself as a similar
hang during configuration of base-files, which is the second package
configured.  strace inside UML produces suspiciously similar results:

write(1, "\n", 1)                       = -1 EPIPE (Broken pipe)
--- SIGPIPE (Broken pipe) ---

The key seems to be that in 0.1.16.4, debootstrap runs "smallyes ''", where
smallyes is a shell function:

smallyes() {
  YES="${1-y}"
  while true ; do echo "$YES" ; done
}

and pipes the output into for dpkg --configure perl-5.005-base.  debootstrap
0.1.17 does the equivalent for all packages.  It appears to be this echo
which I am seeing in strace above.  I would think that after receiving
SIGPIPE, the while loop would be broken, but it doesn't seem to happen that
way, and the process goes on forever.  Indeed, if I replace 'echo' (shell
builtin) with /bin/echo, I get:

/bin/echo: write error: Broken pipe

after each package configuration, and installation continues, though much
more slowly.

This problem goes away if I replace my shell (ash) with bash, so I suspect
this is an ash bug.  I can't seem to reproduce it outside of UML, though.

Any ideas?

-- 
 - mdz